What is Forefront?
Forefront is a platform for developers who want to fine-tune and run open-source language models using their own data. It brings model selection, training, evaluation, dataset management, and inference into one service, with a Playground for testing and APIs for applications. Teams can start with an existing model, store production responses and data, then prepare datasets for fine-tuning as their needs develop; they can also export models for hosting elsewhere.
Forefront Features:
Model Fine-Tuning: Choose an open-source model and customize it with your data for a particular task.
Validation and Evals: Check performance on validation sets and run evaluations such as MMLU, TruthfulQA, MT-Bench, ARC, HumanEval, and AGIEval.
Training Monitoring: View built-in loss charts during training and inspect validation results for a fine-tuned model.
Serverless Inference: Run models through chat or completion API endpoints, or experiment with them in the Playground.
Dataset Workspace: Store training, validation, and evaluation data as datasets ready for fine-tuning, and send production data to Forefront through its API.
Data Inspection: Review dataset samples and explore distributions, including tokens per sample and tokens by label.
Model Portability: Import a model using its Hugging Face model string and export models to host with another provider or on your own.
Forefront Benefits:
Task-Specific Results: Training with examples from a particular task can make a model's output better suited to that work than relying on a general-purpose model alone.
Less Infrastructure Work: Managed inference avoids the need to handle API servers, GPUs, memory errors, dependencies, CUDA, and batching yourself.
Lower Idle Costs: Usage-based inference means you do not pay for GPU capacity sitting unused when there is no traffic.
Clearer Data Boundaries: Keeping requests unlogged and not using customer data to train models gives teams more control over how their information is handled.
Hosting Flexibility: Exporting a model lets teams choose where it runs instead of being limited to Forefront for hosting.
Use Cases:
Email Summaries: Fine-tune a model on email summary examples for a workflow that needs concise versions of longer messages.
Company Enrichment: Prepare a model for company enrichment tasks using relevant training examples and validation data.
Contact Enrichment: Customize a model for contact enrichment work and check its output against a validation set.
Email Hook Writing: Use examples of email hooks to adapt a model for generating opening lines for an email workflow.
Model Evaluation: Evaluate a fine-tuned model on validation samples and benchmarks before connecting it to an application API.